Alibaba Targets NVIDIA's Hopper With Zhenwu M890 AI Chip, Claiming 3x The H20 Performance, 144GB HBM3 & A Roadmap Through 2028

Alibaba has unveiled the Zhenwu M890, its most advanced AI chip to date, marking a significant escalation in China's effort to break free from dependence on NVIDIA hardware for artificial intelligence workloads. The announcement, made alongside the Qwen 3.7-Max large language model, signals that the global AI chip war has entered a new phase where Chinese tech giants are no longer just catching up but are staking claims on performance benchmarks previously dominated by American silicon.

The Zhenwu M890: By the Numbers

The M890 is built on Alibaba's in-house PPU (Parallel Processing Unit) architecture and features a Transformer core engine specifically designed for AI inference workloads. The headline specifications are striking: 0.6 PFLOPs of FP16 compute, putting it in the same conversation as NVIDIA's A100, and Alibaba claims it delivers three times the performance of the Hopper H20, the restricted chip NVIDIA has been selling to Chinese customers under U.S. export controls.

Memory has also received a substantial upgrade. The M890 packs 144 GB of HBM3, a 50 percent increase over the previous generation Zhenwu 810E's 96 GB. Interconnect bandwidth has been boosted to 800 GB/s, up from 700 GB/s on the 810E. The chip supports FP32, FP16, FP8, and FP4 formats, putting its numerical capability set on par with NVIDIA's latest Rubin architecture and Huawei's Ascent 950 series.

But the real story is not the chip alone, it is the ecosystem Alibaba is building around it.

The Panjiu AL128 Supernode: A Full-Stack Play

Alibaba is not selling the M890 as a standalone accelerator. The chip arrives as part of a vertically integrated system called the Panjiu AL128 Supernode Server, which tightly integrates 128 AI accelerators within a single rack. The system also includes the ICN Switch 1.0 interconnect chip offering 25.6 Tb/s of bandwidth with point-to-point latency under 150 nanoseconds, the Yitian Arm-based host CPU, and Panmai series networking cards.

This full-stack approach mirrors the strategy NVIDIA has used with its DGX systems and NVLink interconnect, but with one critical difference: every component is designed and controlled by a single Chinese company, making it immune to the export restrictions that have hamstrung NVIDIA's ability to serve the Chinese market.

Alibaba reports that its T-Head division has shipped approximately 560,000 Zhenwu AI chips to date, serving more than 400 external customers across 20 industries. That is not a prototype number; it is a production scale that suggests real market traction.

The Roadmap Through 2028

Alibaba is not stopping at the M890. The company announced a two-generation roadmap: the Zhenwu V900 arriving in Q3 2027 with a 3x performance boost, 216 GB of memory, and 1,200 GB/s of bandwidth, followed by the Zhenwu J900 in Q3 2028 with further architectural advances.

This kind of long-term roadmap publication is significant. It tells customers and investors that Alibaba is committed to this product line for the long haul, reducing the risk of adoption for enterprises weighing a transition away from NVIDIA.

Qwen 3.7-Max: The Software Side of the Stack

Alongside the hardware, Alibaba launched Qwen 3.7-Max, a large language model optimized for agentic AI workloads. The model is designed for complex reasoning, long-horizon task execution, and multi-agent workflow orchestration. Alibaba claims it can sustain continuous operation for up to 35 hours and manage over 1,000 tool calls without performance degradation, capabilities that position it as a backbone for autonomous AI agent systems.

The model achieves competitive results across coding, general-purpose agent, and multilingual benchmarks, and is optimized for leading agent frameworks. It will be available through Alibaba's Model Studio platform for global developers.

The Geopolitical Calculus

The M890 launch cannot be separated from the broader geopolitical context. U.S. export controls have progressively restricted NVIDIA's ability to sell its most advanced chips to Chinese customers, first with the A100 and H100 bans, then the H800 restrictions, and most recently the H20 limitations. Each restriction has been designed to slow China's AI advancement, but the Zhenwu M890 suggests that the gap is closing faster than policymakers may have anticipated.

If Alibaba's performance claims hold up under independent testing, Chinese AI companies and cloud providers will have a domestic alternative that is no longer a compromise but a genuine competitor. That changes the strategic calculus entirely, because it means export controls may have accelerated the very outcome they were designed to prevent: a self-sufficient Chinese AI hardware ecosystem.

What This Means For You

The Zhenwu M890 is more than a chip announcement; it is a signal that the global AI supply chain is fragmenting into distinct ecosystems. For technology investors, the implication is that NVIDIA's near-monopoly on high-performance AI training and inference hardware faces an increasingly credible challenger in the Chinese market, which represents roughly 20 percent of global AI compute demand. For developers and enterprises building AI systems, Alibaba's full-stack approach means there will soon be a viable alternative to the NVIDIA ecosystem for inference workloads, potentially at lower cost and without the supply chain risks that export controls have introduced. For policymakers, the question is whether restricting chip sales to China has delayed Chinese AI capability or merely redirected investment toward domestic alternatives that are now beginning to rival the restricted products they were designed to replace.